First I’d like to commend the company on an innovative concept. Compared to peers I think you have something innovative compared to other products that monitor circuits without any intelligence. The magic to Sense is the vision that through analytics you’ll be able to break out the the components through electric signature and thus enable a much higher degree of insight than would otherwise be accomplished. In my humble opinion… it simple, innovative and differentiated.
As a brand new Sense owner, I’m going through the same excitement and anxiety as new devices are discovered and then identified by me. It’s been a couple days since install and two devices have been discovered thus far. One I think is accurate and one I think is flagging an impartial load due to its specific almost always on characteristic. In both cases I’ll evaluate, tag and then provide feedback as revelant.
But the process I am going through got me wondering if Sense and the Data Scientists were missing the broader crowd sourcing opportunity. I have seen a lot of topics one training, but haven’t seen anything on our ability to teach. It’s a subtle difference but in the crowd sourcing world it can be huge. I think Sense has an opportunity to make the user community a broad source for information, by giving us rules to provide inputs. I have used this model in the past in corporate America and have never been disappointed with the results when properly structured.
So the opportunity as I see it is as follows: Sense has a engaged community that like me is probably a bit OCD in that we will want to identify and track every single device possible. Heck if I had my way, every non Always On device would be identified and the Always On would be broken out the the 9 or so devices that took 80% of that always on load. But I don’t think I am alone in that desire which is why I believe you have a community that wants to provide the Sense team inputs. You started that with allowing us to identify manufacturer and model information after a device has been discovered, but the real opportunity is for all of us to proved you inputs prior to discovery.
Every one of us has walked around turning devices on and off after installing our first Sense device. We love gaining the insight which is why we bought Sense in the first place. I believe if you provided us rules and the application (web based, app based, email, or other) we would give you far more infomation than you could mine independently, For example if you asked for a clean power signature of a device for five times with manufacturer and model a lot of us would comply in the hopes it would help our personal device discovery as well as others. We could segment the device separately by turning off other circuits and then after running the device the required times, provide Sense the Manufacturer, Model, Sense ID, Start and Stop Times for each run and any additional inputs needed.
Maybe Sense would want more… Maybe less. But the opportunity is lost if you take none. It is not a commitment for identifying those devices, but it does enhance your data scientist abilities to learn and model patterns by providing them an additional input source. Now of course conceptually I have not thought all the components through, but hopefully this will spur some thinking on how to leverage your community a little differently. Crowd sourcing problems can be extremely impactful and that opportunity doesn’t last forever. The window for success is now.
Hi Tom,
Appreciate your sentiments and ditto, sort of. I don’t know enough about RNN-LSTM based classifiers yet, but I believe that’s the model at the core of Sense’s recognition. Based on what I do know, the best “training drill” would be almost the opposite of your example. You would want to exercise every variant of your device cycle many times against every likely combination of other devices in the house. This isn’t about getting a clean signature, but rather finding an invariant pattern in spite of all the other electrical activity.
Here’s one of the better intros to RNN-LSTM based learning:
I think you have made excellent points and I found the article enlightening. My example was definitely the simplicity version for sure. There is so much opportunity for this community to provide inputs and teach the data scientists about our homes and the patterns if they would just provide the mechanism and rules for providing that input.
If the Sense data team asked for a detailed log of what I did for 4 hours, I’d be the crazy person logging down that I opened the fridge at 8:32 AM and make model was blah blah. Then I used the kettle to heat water using the blah blah model electric kettle for x minutes. And then and then and then… I get that the method of device detection may not be ideal for me to train the device myself at this time, but if that is truly the case then we should all be able to help the data scientists themselves learn so that the algorithims improve over time.
So the concept / idea is not flushed out so hopefully the Sense team can build on it or the community can flush out a workable model. But they have our attention so capture it now or lose the window when we bore and move onto something else.
We really appreciate the passion and intellect of the Sense community (I promise I’m not just trying to flatter you ). It’s amazing to have people so willing and dedicated to improving the product and interested in being even more involved in the detection process.
We do want users to be able to contribute to device detection. There are some ways that you can already contribute, that really do help with detection. These include renaming and recategorizing mystery devices. This information feeds back into our system and allows us to better identify devices. You can also mark when a ‘device is not on’, this sends data to our data science team with which to improve detection algorithms off of. Network Device Identification (which you can turn on in settings), gives us the ability to look at network traffic in order to aid in detection (we recently used this information to detect some smart TV models). We also have plans for users to be able to input information about devices that are in their home, but that we haven’t detected yet.
That being said, there are a lot of challenges and nuance around detection of devices that our data science team is working to tackle. Ultimately, a lot of improving detection comes down to having a ton of data on how different devices operate in the context of the home itself. @kevin1 is right in the sense that the best thing to do is to just use your devices!
Thanks for listening, but I don’t fully agree with your assessment that the best way for device detection is to just use your devices. I think it misses the point of the opportunity at hand. I fully believe the best way to improve device detection is to use devices AND for Sense to improve their algorithms. Your data scientists are constantly working on ways to improve this algorithms and I would find it hard to believe that we couldn’t help them.
Regardless of the problem they are currently working on I can’t imagine if they were asked… what do you need to solve this portion of the problem / or improve your algorithm that they wouldn’t have some interesting asks. Then the question changes from how do I figure this out from raw data to what could the very engaged community provide that would change my approach and help improve the algorithms.
Remember this was never about improving my own personal device detection. That is going to happen over time and only through usage and your teams ability to improve the detection routines. This is about Sense using your community to provide differentiated data sets directly to the data scientists that help them solve the challenges they are working on. My personally experience working on data intensive challenges is that when you only work with raw data the data scientists are limited into the insights that can be generated. But… when just the right context is provided to that same data set a previously unsolvable problem becomes clear. That conceptually is what I see as the opportunity for Sense.
Ultimately only Sense can decide what they need to improve and become a better product, but your community is engaged. Why waste the opportunity?
To be clear, I didn’t mean to suggest that work didn’t need to or wasn’t happening on our algorithms. Our data science team is working hard to improve those. I only meant to suggest that the more data they have to work with (based on devices being used), the easier it is for them to do so. Which sounds like what you’re saying as well to some extent.
In the past we’ve run surveys when starting work on new detection algorithms (eg. electric vehicles) and do get a lot of crucial information via renaming/categorization. As I mentioned before, we also have plans for functionality for users to be able to input information about devices that have yet to be detected. We definitely understand and appreciate the interest in being even more involved than that. After all, it would make customers happy and if it helped device detection, would make our team’s job easier. It really just comes down to determining what would be helpful and I assure you that in that regard, it is something we think about. If there is additional user input that our data science team thinks would be effective, we’ll work develop a way to ask for that (be it a survey or an in app feature).
Thanks again for the passion about device detection and Sense as a whole.
I just wonder if it would be helpful for Sense to at least know what devices are in a home. For example, I have a coffee make that, before I reset, Sense had found. Now, nothing. I use it every day. I also have Heats 1-6 which, to me, mean nothing and I can’t map them. If I ‘told’ Sense that I had brand x, model y of a coffee maker it could use that as a filter through the data. Granted, we don’t want data fitting but when Sense hears hoovess it should think horse and not zebras especially when we say we have horses.
If this is already here, I apologize-I haven’t seen it. I, too, am one of those people running around when Sense tells me something turned on to try to figure out what it is, yet I still have number of unidentified devices I can’t seem to figure out. I see people post images of their appliances and their power signatures and people’s feedback as to what they think it is and I love that. However, what about some sort of database of signatures? For instance, if I see an Unnamed Heat identified, I can check the signature mine is showing and scroll though what it might be and that could help me track it down? For instance my Keurig signature is quite common, and I’ve learned that Sense can’t tell it is on, just when it is maintaining its heat, but I know that’s what it is. Thank you for considering, and thanks to everyone that’s posted before for their help to me and everyone else.
For instance my Keurig signature is quite common, and I’ve learned that Sense can’t tell it is on, just when it is maintaining its heat, but I know that’s what it is.
I have noticed the same thing, it only will show the Keurig being on when it is idle, not when it is active and using the most power. I was hoping this would get better with time as I would like to be able to track all the power it is using.
Just posting in this thread because there is such excellent posts! The OPs post is exactly how I feel and wish it was how it worked… If it were possible to just turn off all breakers, turn on one by one, and then manually identify our devices that would be huge… I understand that’s not how sense works at all but we all just want to speed up the process!
I’ve thought the same thing about an archive and was going to start posting some jpgs in an appropriately named thread but haven’t had time to play with it.
I wonder if we could get a signature archive category and have a thread for each main topic. heat, motor, etc…
Just installed my sense last week. My impression this far is thoroughly summed up by the OP. What an enormous opportunity being wasted for many months/years? now… Let us identify our own devices! C’mon sense team…That’s a no-brainer!
Ben (from sense team) seemed to imply back in February that this functionality was in the works. It’s been nearly 5 months. Any progress?
Also agree with poster that suggested users could provide sense scientists work a list of devices in each house to help prevent false positives. Is this possible?
@johnavedian,
I can’t speak on behalf of Sense, but can give some insights on your thoughts in the context of machine learning.
we are already massively crowdsourcing by providing our voltage and current waveforms, along with identification confirmation data to Sense.
some forms of human feedback that seem intuitively valuable may not be, at least today. For instance, unless one is able to repeatedly and accurately mark the start of an on-event or an off-event to the nearest millisecond consistently over a 24 hour period for a series of days, human “training” is likely to be of very limited value. That value may change when the Sense framework is enhanced to identify devices with many-second on and off events plus many minute long steady state periods, the types of events and steady state humans readily identify. Remember, Sense is currently looking for sub-second level on and off signatures, ones you can’t even view in the power meter.
there is room for improvement in Sense land in providing greater structured feedback when identification goes afoul. For instance, the ability to tell Sense that it has conflated two devices, rather than just telling Sense that the device it is proffering is not on.
@kevin1 already gave some of my answers, but I’ll add a bit.
I can’t speak for Ben, as that was before my time, but we have added better functionality to track down your unknowns (I’m guessing he was referring to Community Names). Device detection is also improving constantly, so that should be whittling down the unknowns.
But as Kevin already noted, Sense is at its very core a crowdsourcing project. Your detected devices and the names and labels you give them feedback eventually to the Sense monitors in other users’ homes. I won’t repeat everything he said, but ML is very different than human learning and what you think may provide helpful info intuitively is very likely at odds with how Sense learns (I know that’s been my experience at least).
That said, we can take better user input into account and we’re working on some ways in order to do so. I can’t go into specifics beyond that, but Kevin’s last point is pretty on the mark
I installed my sense about a month ago now, and my excitement over the device is waning. I purchased the device for my small farm with the justification of it being able to track the power usage of some large consumers (and deduct off my farm income tax) - mainly two 5hp irrigation pumps that get used a couple times a day. But nope… Not even my LEAF EV is detected. I have so many electrical devices, yet it has discovered only 16. I would be more than happy to share the specs of these devices and screen shots of the power meter when they turn on. But as I see from this thread, there is no interest in that input data. As a software engineer (that hasn’t done anything with machine learning), I would think that sort of information would be appreciated… but it doesn’t appear so.
I think there is a big market for small farm/business use of this device if it could track the use of business consumption for tax purposes. I am unable to recommend it to anyone for that use at this point because it has so far failed to detect any of my farm related devices. I hope you all continue to improve the produce and maybe someday it will be useful for me. As of now, it is an interesting toy…
Hey Dave. What is your panel setup like? We don’t really advertise small business use, because most small businesses aren’t run on 200A split-phase panels. That’s not a super common spec in a commercial space.
Hi Ryan, it is a 200A panel. split-phase (not 3 phase). Just a small 18acre hobby/family farm. So we have the house and farm all one meter - thus some aspects are a farm expense, others are not. Seems like it would be a common need. I haven’t reached out to support yet.
). It’s amazing to have people so willing and dedicated to improving the product and interested in being even more involved in the detection process.
.